Climate in a Bottle: Towards a Generative Foundation Model for the Kilometer-Scale Global Atmosphere

(By Noah D. Brenowitz, Tao Ge, Akshay Subramaniam, Peter Manshausen, Aayush Gupta, David M. Hall, Morteza Mardani, Arash Vahdat, Karthik Kashinath, Michael S. Pritchard, NVIDIA

* The paper introduces **Climate in a Bottle (cBottle)**, a generative diffusion-based AI framework capable of synthesizing full global atmospheric states at an unprecedented $\mathbf{5 \text{ km resolution}}$ (over 12.5 million pixels per sample). Unlike prevailing auto-regressive paradigms, cBottle samples directly from the full distribution of atmospheric states without requiring a previous time step, thereby avoiding issues like drifts and instabilities inherent to time-stepping models.

* cBottle utilizes a **two-stage cascaded diffusion approach**: a global coarse-resolution generator conditioned on minimal climate-controlling inputs (such as monthly sea surface temperature and solar position), followed by a patch-based 16x super-resolution module.

* The model demonstrates **foundational versatility** by being trained jointly on multiple data modalities, including ERA5 reanalysis and ICON global cloud-resolving simulations. This enables various zero-shot applications such as climate downscaling, channel infilling for missing or corrupted variables, bias correction between datasets, and translation between these modalities.

* cBottle proposes a new form of **interactive climate modeling** through the use of guided diffusion. By training a classifier alongside the generator, users can steer the model to conditionally generate physically plausible **extreme weather events, such as Tropical Cyclones**, at specified locations on demand, circumventing the need to sift through petabytes of output to find rare events.

* The model exhibits **high climate faithfulness** across a battery of tests, including reproducing diurnal-to-seasonal scale variability, large-scale modes of variability (like the Northern Annular Mode), and tropical cyclone statistics. Furthermore, it achieves **extreme distillation** by encapsulating massive datasets into a few GB of neural network weights, offering a 256x compression ratio per channel.

Probabilistic measures afford fair comparisons of AIWP and NWP model output (Tilmann Gneiting, Tobias Biegert, Kristof Kraus, Eva-Maria Walz, Alexander I. Jordan, Sebastian Lerch, June 10, 2025)

Introduction of a New Fair Comparison Metric: The paper introduces the Potential Continuous Ranked Probability Score (PC), a new measure designed to allow fair and meaningful comparisons between single-valued output from data-driven Artificial Intelligence based Weather Prediction (AIWP) models and physics-based Numerical Weather Prediction (NWP) models. This approach addresses concerns that traditional loss functions (like RMSE) may unfairly favor AIWP models, which often optimize their training using these metrics.

 Methodology Based on Probabilistic Postprocessing: PC is calculated by applying the same statistical postprocessing technique—specifically Isotonic Distributional Regression (IDR), also known as Easy Uncertainty Quantification (EasyUQ)—to the deterministic output of both AIWP and NWP models. PC is then defined as the mean Continuous Ranked Probability Score (CRPS) of these newly generated probabilistic forecasts.

 Measure of Potential Skill and Invariance: PC quantifies potential predictive performance. A key property of PC is that it is invariant under strictly increasing transformations of the model output, treating both forecasts equally and facilitating comparisons where the pre-specification of a loss function might otherwise place competitors on unequal footings.

 AIWP Outperformance and Operational Proxy: When applied to WeatherBench 2 data, the PC measure demonstrated that the data-driven GraphCast model outperforms the leading physics-based ECMWF high-resolution (HRES) model. Furthermore, the PC measure for the HRES model was found to align exceptionally well with the mean CRPS of the operational ECMWF ensemble, confirming that PC serves as a reliable proxy for the performance of real-time operational probabilistic products.

Jigsaw: Training Multi-Billion-Parameter AI Weather Models With Optimized Model ParallelismAuthors: Deifilia Kieckhefen, Markus Götz, Lars H. Heyen, Achim Streit, and Charlotte Debus (Karlsruhe Institute of Technology, Helmholtz AI)

The paper introduces WeatherMixer (WM), a multi-layer perceptron (MLP)-based architecture designed for atmospheric forecasting, which serves as a competitive alternative to Transformer-based models. WM's workload scales linearly with input size, addressing the scaling challenges and quadratic computational complexity associated with the self-attention mechanism in Transformers when dealing with gigabyte-sized atmospheric data.• A novel parallelization scheme called Jigsaw parallelism is proposed, combining both domain parallelism and tensor parallelism to efficiently train multi-billion-parameter models. Jigsaw is optimized for large input data by fully sharding the data, model parameters, and optimizer states across devices, eliminating memory redundancy.

 Jigsaw effectively mitigates hardware bottlenecks, particularly I/O-bandwidth limitations frequently encountered in training large scientific AI models. Due to its partitioned data loading (domain parallelism), the scheme achieves superscalar weak scaling in I/O-bandwidth-limited systems.

 The method demonstrates excellent scaling behavior on high-performance computing systems, exceeding state-of-the-art performance in strong scaling in computation–communication-limited systems. The training was successfully scaled up to 256 GPUs, reaching peak performances of 9 and 11 PFLOPs.• Beyond hardware efficiency, Jigsaw improves predictive performance: by partitioning the model across more GPUs (model parallelism) instead of relying solely on data parallelism, it naturally enforces smaller global batch sizes, which empirically helps mitigate the problematic large-batch effects observed in AI weather models, leading to lower loss values.

XiChen: An observation-scalable fully AI-driven global weather forecasting system with 4D variational knowledgeAuthors: Wuxin Wang, Weicheng Ni, Lilan Huang, Tao Han, Ben Fei, Shuo Ma, Taikang Yuan, Yanlai Zhao, Kefeng Deng, Xiaoyong Li, Boheng Duan, Lei Bai, Kaijun Ren

XiChen is the first observation-scalable fully AI-driven global weather forecasting system. Its entire pipeline, from Data Assimilation (DA) to 10-day medium-range forecasting, can be accomplished within only 17 seconds using a single A100 GPU. This speed represents an acceleration exceeding 400-fold compared to the computational time required by operational Numerical Weather Prediction (NWP) systems.

 The system is architected upon a foundation model that is initially pre-trained for weather forecasting and subsequently fine-tuned to function as both observation operators and DA models. Crucially, the integration of four-dimensional variational (4DVar) knowledge ensures that XiChen’s DA and medium-range forecasting accuracy rivals that of operational NWP systems.

 XiChen demonstrates high scalability and robustness by employing a cascaded sequential DA framework to effectively assimilate both conventional observations (GDAS prepbufr) and raw satellite observations (AMSU-A and MHS). This design allows for the future integration of new observations simply by fine-tuning the respective observation operators and DA model components, which is critical for operational deployment.

 In terms of performance, XiChen achieves a skillful weather forecasting lead time exceeding 8.25 days (with ACC of Z500 > 0.6). This result is comparable to the Global Forecasting System (GFS) and substantially surpasses the performance of other end-to-end AI-based global weather forecasting systems, such as Aardvark (less than 8 days) and GraphDOP (about 5 days).

 A dual DA framework is implemented to operationalize XiChen as a continuous forecasting system. This framework utilizes separate 12-hour and 3-hour Data Assimilation Windows (DAW) to circumvent the multi-hour latency characteristic of high-resolution systems (like IFS HRES), thereby enabling the real-time acquisition of medium-range forecast products.

A data-to-forecast machine learning system for global weather

Xiuyu Sun et al. (2025). A data-to-forecast machine learning system for global weather. Nature Communications, https://doi.org/10.1038/s41467-025-62024-1

• FuXi Weather is introduced as a groundbreaking end-to-end machine learning system for global weather forecasting. It autonomously performs data assimilation and forecasting in a 6-hour cycle, directly processing raw multi-satellite observations, and notably, it is the first such system to demonstrate continuous cycling operation over a full one-year period.

• The system exhibits superior forecast accuracy in observation-sparse regions, outperforming traditional high-resolution forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF HRES) beyond day one in areas like central Africa and northern South America, despite utilizing substantially fewer observations

.• Globally, FuXi Weather delivers comparable 10-day forecast performance to ECMWF HRES, generating reliable forecasts at a 0.25° resolution and extending the skillful lead times for a number of key meteorological variables

.• FuXi Weather offers a cost-effective and physically consistent alternative to traditional Numerical Weather Prediction (NWP) systems. Its computational efficiency and reduced complexity are valuable for improving operational forecasts and enhancing climate resilience in regions with limited land-based observational infrastructure

.• This development challenges the prevailing view that standalone machine learning-based weather forecasting systems are not viable for operational use, demonstrating a significant step forward in the application of AI to real-world weather prediction.

Probabilistic Emulation of a Global Climate Model with Spherical DYffusion

Salva Rühling Cachay, Brian Henn, Oliver Watt-Meyer, Christopher S. Bretherton, Rose Yu

• This paper introduces Spherical DYffusion, the first conditional generative model designed for the probabilistic emulation of a global climate model. It achieves accurate and physically consistent global climate ensemble simulations by combining the dynamics-informed diffusion framework (DYffusion) with the Spherical Fourier Neural Operator (SFNO) architecture.

• The model demonstrates significant improvements in climate model emulation, achieving near gold-standard performance. It substantially reduces climate biases compared to existing baselines, with errors often closer to the reference simulation’s noise floor. For example, it reduces climate biases to within 50% of the reference model, outperforming the next best baseline (ACE) by more than 2x.

• Spherical DYffusion enables stable and efficient long-term climate simulations, capable of 100-year simulations at 6-hourly timesteps with low computational overhead. It offers significant speed-ups (over 25x) and energy savings compared to the physics-based FV3GFS model it emulates.

• The method is particularly effective for ensemble climate simulations, accurately reproducing climate variability consistent with the reference model and further reducing climate biases through ensemble-averaging. The paper also highlights that short-term weather performance does not necessarily translate to accurate long-term climate statistics.

FourCastNet 3: A geometric approach to probabilistic machine-learning weather forecasting at scale Boris Bonev, Thorsten Kurth, Ankur Mahesh, Mauro Bisson, Jean Kossaifi, Karthik Kashinath, Anima Anandkumar, William D. Collins, Michael S. Pritchard, and Alexander Keller

• FourCastNet 3 (FCN3) introduces a pioneering geometric machine learning approach for probabilistic ensemble weather forecasting. It is designed to respect spherical geometry and accurately model the spatially correlated probabilistic nature of weather, resulting in stable spectra and realistic dynamics across multiple scales. The architecture is a purely convolutional neural network tailored for spherical geometry.

• Achieves superior forecasting accuracy and speed, surpassing leading conventional ensemble models and rivaling the best diffusion-based ML methods. FCN3 produces forecasts 8 to 60 times faster than these approaches; for instance, a 60-day global forecast at 0.25°, 6-hourly resolution is generated in under 4 minutes on a single GPU

.• Demonstrates exceptional physical fidelity and long-term stability, maintaining excellent probabilistic calibration and realistic spectra even at extended lead times of up to 60 days. This crucial achievement mitigates issues like blurring and the build-up of small-scale noise, which challenge other machine learning models, paving the way for physically faithful data-driven probabilistic weather models.

• Enables scalable and efficient operations through a novel training paradigm that combines model- and data-parallelism, allowing large-scale training on 1024 GPUs and more. All key components, including training and inference code, are fully open-source, providing transparent and reproducible tools for meteorological forecasting and atmospheric science research.

Can AI weather models predict out-of-distribution gray swan tropical cyclones?by Y. Qiang Sun, Pedram Hassanzadeh, Mohsen Zand, Ashesh Chattopadhyay, Jonathan Weare, and Dorian S. Abbot

• Inability to Extrapolate to Gray Swans Globally: AI weather models like FourCastNet struggle to predict "gray swan" tropical cyclones (TCs), which are rare, strong, and absent from training data. When Category 3-5 TCs are entirely removed from the global training dataset, the model cannot extrapolate from weaker storms (Category 1-2) to accurately forecast these stronger, unseen events, often leading to dangerous "false negative" predictions. This limitation persists even if the training data includes strong extratropical cyclones, as their dynamics differ from TCs.

• Limited Generalization Across Basins for Dynamically Similar Events: Despite the global extrapolation challenge, FourCastNet can demonstrate some ability to generalize learning across tropical basins for dynamically similar strong storms. This means that if the model has seen strong TCs in one ocean basin, it can apply that learned knowledge to forecast similar strong TCs in another basin, even if those specific events were excluded from the training data for that particular region.

• Lack of Physical Consistency and Masked Performance: Current AI weather models, including FourCastNet, fail to reproduce key physical balances like the gradient-wind balance that TCs obey in real-world data, regardless of whether they were trained on full or reduced datasets. Furthermore, common evaluation metrics (e.g., anomaly correlation coefficient or root-mean-square error) can obscure these critical shortcomings by showing similar overall performance for general weather or less extreme events, highlighting the need for specialized tests for gray swans.

• Implications and Future Directions: This research suggests that current AI weather models may provide unreliable early warnings for unprecedented extreme weather events, potentially leading to serious societal risks. It also indicates that AI climate emulators might mischaracterize extreme weather statistics for gray swans. The study emphasizes the urgent need for novel learning strategies (such as incorporating physics-based synthetic data or rare-event sampling algorithms) and rigorous testing methodologies to improve and reliably validate AI models for these high-impact, out-of-distribution events.

Probabilistic Emulation of a Global Climate Model with Spherical DYffusionby Salva Rühling Cachay, Brian Henn, Oliver Watt-Meyer, Christopher S. Bretherton, Rose Yu

• The paper introduces Spherical DYffusion, the first conditional generative model for probabilistic emulation of a realistic global climate model, offering efficient and accurate climate ensemble simulations.
• It demonstrates that weather forecasting performance is not a strong indicator of long-term climate performance, a crucial insight for developing climate models.
• Spherical DYffusion significantly reduces climate biases compared to existing baselines like ACE and DYffusion, achieving errors often closer to the reference simulation's "noise floor".
• The model generates stable, 10-year-long probabilistic predictions with minimal computational overhead, being more than 25 times faster than the physics-based FV3GFS model it emulates, while also reproducing consistent climate variability.

"Do AI models produce better weather forecasts than physics-based models? A quantitative evaluation case study of Storm Ciarán"

By Andrew J. Charlton-Perez, Helen F. Dacre, Simon Driscoll, Suzanne L. Gray, Ben Harvey, Natalie J. Harvey, Kieran M. R. Hunt, Robert W. Lee, Ranjini Swaminathan, Remy Vandaele & Ambrogio Volonté. Published in partnership with CECCR at King Abdulaziz University, Nature,

DOI: 10.1038/s41612-024-00638-w.

Here are the main takeaways from the paper:• AI models (FourCastNet, Pangu-Weather, GraphCast, FourCastNet-v2) demonstrate strong capabilities in capturing large-scale dynamical drivers vital for rapid storm development, such as the storm's position relative to upper-level jets. They also accurately reproduce the larger synoptic-scale structure of cyclones like Storm Ciarán, including the cloud head's position and the warm sector's shape.

Despite these strengths, AI models consistently underestimate the peak amplitude of winds, both at the surface and in the free atmosphere, associated with storms. They also struggle to resolve detailed structures crucial for issuing severe weather warnings, such as sharp bent-back warm frontal gradients, and show variable success in capturing warm core seclusion.

The underestimation of strong winds is not a consequence of the AI models' output resolution or their training data. This discrepancy persists even when compared against ERA5 (on which these models were trained) and numerical weather prediction (NWP) models of similar resolution, suggesting a more fundamental limitation in their ability to represent intense wind features.

The case study of Storm Ciarán highlights the pressing need for a more comprehensive assessment of machine learning weather forecasts. Moving beyond isolated error metrics to evaluate all relevant spatio-temporal features of physical phenomena is essential for identifying specific areas for improvement and fostering rapid advancements in this new and potentially transformative forecasting tool.